DeepTeam
1,566Open-source LLM red-teaming framework from Confident AI with 40+ attack vulnerabilities.
Maturity tier details
Computed tier: 2B(raw score 4.28). Effective tier: 2B.
Facets:documented_in=0evaluated_on=0integration_surfaces=1paper_citations=0recent_commit=2stars=1.28
Source of truth:pipeline/maturity/score.py
Description
Open-source red-teaming framework for LLM applications, from the authors of DeepEval. Ships a library of vulnerabilities (prompt injection, jailbreak, PII leakage, bias, toxicity, misinformation, excessive agency, unauthorized actions) and attack enhancement strategies (multi-turn, roleplay, encoding, linguistic obfuscation) that stress-test an LLM under test via programmatic attack agents. Findings are emitted as scored reports suitable for pre-deployment gating and continuous-evaluation harnesses.
Why it's included
Occupies detection × post_generation × fully_autonomous with a red-team-specialist locus that no other modeled node covers. Adversarial pre-deployment scanning is analytically distinct from trace-capture observability — it proactively synthesizes attack inputs rather than sampling production traffic — and maps cleanly onto the security_vulnerability and supply_chain_attack failure modes.
Security notes
Open-source (Apache-2.0). The framework generates adversarial prompts at scale — intended strictly for systems under test. Calls the model-under-test plus a judge model; budget token spend accordingly. Attack corpora evolve with upstream releases; pin versions for reproducible audit evidence.
Addresses failure modes
- Security vulnerability — Insecure patterns, injection flaws, weak cryptography, secret leakage.
- Supply-chain attack — Slopsquatting, typosquatting, compromised MCP servers.
- Fabrication — Generating reference to non-existent APIs, packages, or symbols.